How Veribix measures

Two pillars, kept separate

Veribix measures two different things and never mixes them into one number. The first is first-party truth: what AI assistants actually sent to your site, read from your own analytics, your Search Console, your server logs and, where you proxy, your Cloudflare data. The second is answer measurement: whether assistants name your company when someone asks a question in your category.

They are reported separately because they can disagree, and the disagreement is usually the most informative thing available. A company named constantly that receives nobody has a click-through problem. A company receiving traffic that converts badly has a landing page problem. Collapsing both pillars into a single visibility score hides which of the two you have.

How answers are sampled

A model asked the same question twice returns different companies. That is a property of the systems being measured rather than an error, and it determines the only honest way to report presence.

Veribix runs each question repeatedly and reports the share of runs in which a company appeared, with the number of runs shown beside it. Eleven of twenty is a complete statement; fourth place is not, because it implies an ordering that does not survive asking again.

Questions fall into two kinds. Category questions contain no company name and are what reveal whether you are in the consideration set at all; these run once against a shared corpus and serve every customer in that category, which is what makes a free audit affordable. Branded questions contain your name or a competitor's and are private to your account.

Measurement uses provider APIs. These are a reproducible proxy for the consumer applications rather than the applications themselves, and where a setting such as browsing materially changes an answer, the setting is recorded with the result.

Evidence, and why every figure links to it

Every stored answer is written to an append-only log with a hash chain, so a record cannot be edited or removed without the chain breaking. A verification command re-computes the chain and reports any gap.

This exists because the category's central problem is that nobody can check anything. A visibility score with no underlying answers is a claim. A share of runs with the twenty stored answers behind it is evidence, and you can read them.

Figures derived from your own systems carry the same treatment in a different form: a crawler count resolves to redacted sample log lines, and a session figure resolves to the analytics rows it was read from.

Confidence labelling

Every referral figure carries one of two labels. Measured means it was counted from data that identified its source. Estimated means it was inferred, and the reason is stated.

The most common reason for an estimate is referrer stripping. Several assistants send outbound links with the referring page removed, so their visits arrive labelled as direct traffic in analytics and in server logs alike. Nothing recovers that connection afterwards, for anyone.

So an identified AI referral figure is a floor rather than a total, and Veribix says so beside the number rather than in a footnote. A figure whose limits are stated next to it is one people trust; a figure whose limits emerge later under questioning is one they stop trusting entirely.

Redaction at ingestion

Server logs contain personal information inside URLs. Veribix redacts at the point of ingestion rather than afterwards: query parameters are dropped unless they are campaign or source tags, and IP addresses are used only to separate automated crawlers from human visitors and are not retained in the extracted data.

Raw uploaded logs are deleted after extraction. What is kept is the aggregate and a small number of redacted sample lines that let a customer verify a figure by hand.

If a file contains a URL that cannot be redacted reliably — an address embedded in a path, for instance — the file is rejected and the reason given, rather than being partially ingested. Silently dropping unreadable lines would understate a figure without saying so, which is the worse of the two failures.

What Veribix does not claim to measure

This section matters more than the rest and it is the one most of this category does not publish.

Veribix cannot measure a mention that produced no click. If an assistant names you and the reader simply remembers it, that influence is real and no tool detects it.

Veribix cannot attribute a visit whose referring page was never sent. That traffic sits in your direct bucket mixed with people who typed your address, and no configuration, vendor or server-side setup separates them, because the information never left the browser.

Veribix cannot tell you that a change you made caused a change in your visibility. It can report that both happened and whether the difference exceeds the variation the measurement normally shows. Where no effect is detectable, it says so — that is a first-class result rather than a failed one, and it is far more common than published case studies suggest.

Veribix does not report a position, a place, or any single score for presence in AI answers, because no stable ordering exists to report. It does not write content, and it does not make changes to your site.